Extracting Basic Fighter Maneuvers from Actual Flight Data Mustafa Karli, Mehmet Önder Efe, and Hayri Sever Department of Computer Engineering, Hacettepe University, Beytepe Ankara, 06800, Turkey Email: mustafa.karli@hacettepe.edu.tr, {onderefe, hayri.sever}@gmail.com Abstract—Air combat maneuvers are very complex actions performed by agile aircrafts. Extracting critical maneuvers from a combat scenario in a structured format has many advantages like teaching maneuvers to the unmanned systems, evaluating pilot performance or analyzing possible combat scenarios. Basic Fighter Maneuvers are special maneuvers that are building blocks of combat fighting. This article proposes a methodology to identify pre-defined movements, match well-known combat maneuvers in a real flight of agile combat aircraft and build a feasible corpus to use this data for machine learning. The claims of the paper are justified by the simulation results. Index Terms—flight parsing, basic fighter maneuvers, air combat I. INTRODUCTION Basic Fighter Maneuvers are executed by agile aircraft during "Within Visual Range" in defensive or offensive positions or missile evacuation. For training artificial systems like UAVs, besides domain information [1], one of the best learning sources is real flight information of manned air vehicles. There are auto-pilot designs and combat support systems in the literature like rule based systems [2], influence diagrams [3], human cognitive models [4] and Artificial Intelligence (AI) techniques for air combat maneuvering [5]. There is a comparison of artificial neural networks and rule based system in [6] and maneuver prediction in [7] to support human combat pilots. Autonomous control of UAV is designed using ANFIS in [8]. There is an additional design by ANFIS in [9] including a predefined flight path. Both ANFIS design is for a single UAV without combat fighting. Assuming air combat as a pursuer-evader game and optimizing using approximation and dynamic programming is presented in [10]. Composing a flight trajectory in terms of seven primitive actions and a way point decomposition algorithm is presented in [11]. There is also a sliding mode controller design by the same author proposed in [12] which excludes an arbitrary movement mode. Our work advances the subject area in terms of representing a maneuver by movement sections instead of many flight parameters and proposes an abstraction stack for flight representation. The real flight data of the agile Manuscript received February 6, 2017; revised June 28, 2017. aircrafts are decomposed into meaningful movement sequences and BFM maneuvers are searched and labeled to be learned by machine learning systems. This paper is organized as follows: In the next section we introduce the definitive terms of the problem. The third section proposes an abstraction stack for air frame flight representation. Using this abstraction, air operations can be executed from mission planning to physical control layer. The forth section defines the basic fighter maneuvers of close air maneuvers in terms of proposed abstraction. The fifth section defines how real flight information and relative geometry of two fighting aircrafts are decomposed. The sixth section evaluates methods for searching and indexing BFM in flight data and proposes a specific search method. The seventh section discusses the benefits of the proposed approach with simulation results and evaluates the search method. The last section defines the required steps to be performed for machine learning techniques with the concluding remarks. II. PROBLEM DEFINITION The objective of an air combat scenario is to move the aircraft into a position where one can shoot the other aircraft or minimize the risk of being shot. This depends on the positional advantage of both aircrafts which depends on the “relative geometry” to each other. Human pilot control the aircraft using the stick and gas pedal where a series of physical, aerodynamic and atmospheric equations run through propulsion, ailerons, elevators, rudder, wing and platform surface resulting forces and accelerations on 3 dimensions which changes the state of the system. This is a non-linear system control that is also affected by non-deterministic conditions like atmosphere, gravitational changes, varying weight and center of gravity. The air frame has 12 state variables  = {, , ℎ, , , , , , , , , } which are north, east, height position, velocity, roll, pitch, heading body axes angles, angular difference of pitch and heading between body and velocity axes and body angular velocity in three dimensions. Air combat is using an aircraft as a weapon and has its own domain rules to learn and practice. In a real air combat, both sides are maneuvering instantly to take advantage. Both sides can be in offensive position while was defensive in the previous action of the engagement. So classical pursuer-evader tactics is not applicable since pursuer only considers pursuing and evader only considers Lecture Notes on Information Theory Vol. 5, No. 1, June 2017 1 ©2017 Lecture Notes on Information Theory doi: 10.18178/lnit.5.1.1-6